Optimised TokenText (fat ptr)

[iDawer]

Following https://github.com/rust-analyzer/rust-analyzer/pull/8209 Basically it is a fat pointer Nonnull<str> with owning semantics, and uses field offset to get rid of indirection. The question is: It's not that mad, is it? I need someone to agree with me so I could have finished the implementation. :smile:

[matklad]

Yeah, I think this belongs here! Although let's start with a double-indirection safe implementation firs. I won't expect this to be a bottleneck, and I would expect a lot of churn in rowan.

But I do think we should have something along this lines long-term. We can even do a crazier thing, and borrow nodes of immutable syntax trees. Hm, now that I've spelled this out, I wonder if we actually should just return Cow<str, String> from that rust-analyzer method? This probably requires exposing is_mutable getter...

[iDawer]

I'll open a new PR with with the rust-analyzer's implementation

[iDawer]

I'll open a new PR with with the rust-analyzer's implementation

nvm

We can even do a crazier thing, and borrow nodes of immutable syntax trees. Hm, now that I've spelled this out, I wonder if we actually should just return Cow<str, String> from that rust-analyzer method?

Borrowing nodes of immutable syntax trees means NodeData.mutable once created never changes, And mixing immutable syntax nodes with mutable ones is not allowed. Are these invariants hold in code? This really intrigued me. I would investigate on it before moving ra's TokenText.

[matklad]

NodeData.mutable once created never changes, And mixing immutable syntax nodes with mutable ones is not allowed.

Yup, the relevant place to document that would be: https://github.com/rust-analyzer/rowan/blob/clone-for-update/src/cursor.rs#L17-L75.

Ohhhhhh. Oh wow, we can actually change the green method in rowan itself to return a Cow<&GreenNode, GreenNode>. Sweeeeet!